AI Coding Interviews Now Test Judgment, Not Syntax
Western engineering orgs adopting AI-assisted workflows face the same filtering problem: how to identify developers who can steer AI output rather than just accept it. The interview techniques described here — live AI tool use, "when did you veto AI's code," and type-safety for nondeterministic outputs — are directly portable to any team hiring for AI-augmented roles.
A wave of Chinese frontend interview reports shows a decisive shift: candidates are told to use Cursor or Claude Code during the interview, and the real test is whether they can spot the flaws AI bakes into its first draft. Questions center on SSE streaming lag, type-safety for unpredictable model outputs, and discriminated unions for agent state — none of which AI generates correctly by default. Interviewers are filtering out candidates who paste AI output without review, labeling them "passengers" rather than "drivers."
A parallel discussion from LeadDev flags "Agentic Inequality" — the compounding advantage granted to developers who can afford premium tools like Claude Code Max. An intern who answered "none" when asked about Claude Code skills wasn't necessarily unskilled; they were priced out of practice. The barrier is real, but the article argues free alternatives like DeepSeek's open-source Harness and its cheap API let anyone build the same judgment skills without a $200/month subscription.
The core competency being measured is no longer writing code, but vetoing AI-generated code with engineering rationale. Candidates who can explain why they rejected an AI suggestion, or who proactively add AbortController and useDeferredValue to a naive streaming implementation, signal the driver mentality that separates hires from rejections.
Live AI-tool interviews invert the traditional coding test: the artifact matters less than the candidate's critique of it, making code review skill the primary signal.
The "driver vs. passenger" framing captures a genuine hiring heuristic — developers who treat AI output as a first draft to be hardened are qualitatively different from those who treat it as a finished answer.
Agentic Inequality is a useful lens: when a tool acts as labor rather than a passive instrument, cost disparities compound because practice hours translate directly into skill gaps that interviews measure.
Free-tier tools and open-source agent harnesses can close the practice gap, but only if candidates recognize that depth of use — not subscription tier — is what interviews actually probe.
The hardest-to-fake interview signal is explaining the principles behind AI-generated code, because it requires understanding the problem domain independently of the tool that produced the solution.